Method for evaluating the relative state of an aircraft engine
The method addresses the challenge of monitoring aircraft engine aging by training an embedding function and classifier on time series data to evaluate engine state, effectively accounting for aging and maintenance intervals, thus improving engine condition assessments and maintenance timing.
Patent Information
- Application Number
- EP2021740146
- Authority / Receiving Office
- EP · EP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2020-07-06
- Filing Date
- 2021-06-21
- Publication Date
- 2024-11-13
- Estimated Expiration
- 2041-06-21
AI Technical Summary
Current methods for monitoring aircraft engine state do not effectively account for engine aging over time without requiring prior knowledge or expertise, and fail to explicitly use temporal information to predict improvements in engine condition.
A method that creates datasets from time series data recorded during flights, trains an embedding function and classifier to learn temporal order, and calculates an engine state indicator by minimizing a cost function, allowing for the explicit consideration of engine aging and maintenance intervals.
Enables the evaluation of aircraft engine state by accounting for aging without prior knowledge or modeling, preventing false improvements in engine condition assessments and triggering maintenance alarms when necessary.
Smart Images

Figure IMGF0001 
Figure IMGF0002 
Figure IMGF0003
Abstract
Description
DOMAINE TECHNIQUE DE L'INVENTION
[0001] The technical field of the invention is that of monitoring the condition of aircraft engines.
[0002] The present invention relates to a method for assessing the condition of an aircraft engine, and more particularly to a method for assessing the relative condition of an aircraft engine. The present invention also relates to a method for monitoring the relative condition of an aircraft engine. ARRIERE-PLAN TECHNOLOGIQUE DE L'INVENTION
[0003] To anticipate aircraft engine failures, it is known to monitor the engine's condition using data recorded during successive aircraft flights, which evolves over time. Several families of methods are commonly used for this purpose.
[0004] One family of methods involves selecting a set of physical and / or mechanical indicators representative of engine efficiency from the available data and monitoring the values of these indicators over time. However, the choice of this set of indicators relies on industry expertise, particularly knowledge of the data and its values in different flight contexts. Furthermore, the set of indicators often includes a significant amount of data that may not be available for all flights.
[0005] A second family of methods involves using a physical or statistical model, fitting it to the data, and then monitoring indicators, such as model parameters or residuals. However, such methods require prior knowledge of the engine's physical model or the construction of a surrogate statistical model that explains the data dynamics.
[0006] A third family of methods involves tracking hidden variables extracted from the data by decomposition into explanatory factors.
[0007] Examples of methods for monitoring the operation of an aircraft engine are disclosed by documents FR3006785 A1, FR2939170 A1 and WO2020115440 A1.
[0008] Regardless of the method used, the values of the monitored indicators or variables are obtained for each flight, and then a trend is extracted from the evolution of these values across all flights. Thus, these methods do not explicitly use temporal information, such as the fact that the aircraft ages over time, and nothing forces the monotony of the calculated indicators or variables over time, which could predict an improvement in engine condition over certain periods, even when no maintenance has been performed.
[0009] There is therefore a need to assess the condition of an aircraft engine by detecting an aging signal in data recorded during successive flights of the aircraft available for all flights and evolving over time, without using prior knowledge. RESUME DE L'INVENTION
[0010] The invention offers a solution to the problems mentioned above, by making it possible to assess the condition of an aircraft engine by taking into account the aging of the engine over time, from data classically recorded during successive flights of the aircraft, without requiring professional expertise.
[0011] A first aspect of the invention relates to a method for evaluating the relative state of an aircraft engine for a given flight of the aircraft, the aircraft having carried out prior to the given flight a plurality of maintenance operations and a plurality of flights between two successive maintenance operations, each maintenance operation being associated with a maintenance time, a plurality of sets of variables being recorded during each flight in a recording order, each flight being associated with a time series corresponding to a flight time and comprising the plurality of sets of variables recorded during the flight ordered according to the recording order, the method comprising the following steps: For each maintenance, a dataset is created containing the time series associated with each flight performed between the maintenance time associated with the current maintenance and the maintenance time associated with the previous maintenance; a set of datasets is created, including each created dataset, and the set of datasets is separated into a training set and a validation set; joint training of an embedding function and a classifier by minimizing a cost function, including the following substeps: For each dataset in the training set: a pair of time series is randomly selected, each consisting of a first time series and a second time series; the embedding function is applied to the pair of time series to obtain a pair of embeddings;Using the classifier on the pair of dives to obtain a probability that the first time series corresponds to a flight time prior to the flight time corresponding to the second time series; Calculating the cost function from the probability obtained, the flight time corresponding to the first time series and the flight time corresponding to the second time series; Optimizing the dive function and the classifier by minimizing the calculated cost function; Selecting a reference time series from among the time series included in the validation set, including a substep of applying the trained dive function to the reference time series to obtain a reference dive;Calculation of an engine status indicator for the given flight, comprising the following sub-steps: Application of the trained dive function to the time series associated with the flight, to obtain a dive; Calculation of a distance between the dive and the reference dive, the calculated distance corresponding to the engine status indicator.
[0012] Thanks to the invention, a time series containing data recorded successively during the flight, ordered according to time, is constructed for each flight. Time series corresponding to flights that took place between two identical maintenance periods are grouped into a single dataset.
[0013] The invention is based on the assumption that information about engine aging is contained in the data and that aging is a monotonic process with respect to time.
[0014] To extract a monotonic signal reflecting aging, an embedding function is trained on each dataset from a set of training datasets using a classifier that forces the embedding function to learn to place, from their time series, each flight corresponding to the same dataset in temporal order in an embedding space.
[0015] Once trained, the plunge function can provide a representation of the various time series recorded throughout the aircraft's life, ordered chronologically within the plunge space. Calculating the distance within the plunge space between the flight in question and a previous flight performed by the aircraft, chosen as the reference flight, then yields an engine condition indicator that takes aging into account.
[0016] Because the dive function was trained by data set, the indicator is monotonic between two engine maintenances, which prevents the evaluated engine condition from improving without performing maintenance.
[0017] The invention therefore makes it possible to evaluate the condition of the engine by explicitly taking into account aging and without requiring engine modeling or time series, unlike the prior art.
[0018] In addition to the characteristics mentioned in the preceding paragraph, the evaluation method according to the first aspect of the invention may have one or more complementary characteristics from among the following, considered individually or according to all technically possible combinations.
[0019] According to one embodiment, the step of selecting the reference time series comprises the following sub-steps: Applying the trained embedding function to each time series in the validation set, to obtain an embedding for each time series in the validation set; For each time series in the validation set: Calculating the distance between the embedding of the time series and the embedding of each other time series in the validation set; For each dataset in the validation set, creating an aging curve associating each flight instant associated with a time series in the validation set included in the dataset, with the corresponding calculated distance; For each aging curve, calculating a monotonicity indicator; Calculating an aging indicator from the calculated monotonicity indicators; the reference time series being the time series for which the calculated aging indicator meets an aging condition.
[0020] According to an alternative embodiment of the preceding embodiment, the substep for calculating the monotonicity indicator comprises the following substeps: Initialize the monotony indicator to zero; For each time interval between two successive flight times, calculate the first-order time derivative of the aging curve: If the derivative is positive, increment the monotony indicator by 1.
[0021] According to a sub-variant of the previous embodiment, the aging indicator is the sum of the calculated monotonicity indicators.
[0022] According to an alternative implementation of the previous sub-variant implementation, the aging condition is verified for the maximum calculated aging indicator.
[0023] Thus, the selected reference flight is the flight associated with a time series of the validation set corresponding to the healthiest possible engine state, i.e. the flight allowing for the most monotonic aging curves possible.
[0024] According to an alternative embodiment compatible with the previous embodiment, the evaluation method according to the first aspect of the invention further includes a step of triggering an alarm if the engine status indicator is above an alarm threshold.
[0025] According to a sub-embodiment of the previous embodiment, the alarm threshold is equal to the maximum aging indicator preceding maintenance within a predefined time margin.
[0026] Thus, an alarm is triggered when the assessed condition of the engine has reached a level of aging beyond which maintenance should be performed to anticipate engine failure. The longer the time buffer, the lower the risk of engine failure occurring before the next maintenance.
[0027] A second aspect of the invention relates to a method for monitoring the relative state of an aircraft engine over a plurality of flights comprising the steps of the evaluation method according to any one of the preceding claims for one flight of the plurality of flights, and the evaluation step for each other flight of the plurality of flights.
[0028] A third aspect of the invention relates to a computer program product comprising instructions which, when the program is executed on a computer, lead the computer to implement the steps of the evaluation process according to the first aspect of the invention or the steps of the monitoring process according to the second aspect of the invention.
[0029] The invention and its various applications will be better understood by reading the following description and examining the accompanying figures. BREVE DESCRIPTION DES FIGURES
[0030] The figures are presented for illustrative purposes only and are in no way limiting to the invention. There figure 1 is a synoptic diagram illustrating the sequence of steps in a method for evaluating the condition of an aircraft engine according to the first aspect of the invention. figure 2 is a synoptic diagram illustrating the sequence of substeps for a first embodiment of the fourth step of the process according to the first aspect of the invention. figure 3 This illustrates an example of the flights and maintenance performed by the aircraft over time. figure 4 shows a schematic representation of the third step of the process according to the first aspect of the invention. figure 5 shows a schematic representation of the fourth and fifth steps of the process according to the first aspect of the invention. figure 6 shows a schematic representation of the first, second, and third substeps of the first embodiment of the fourth step of the process according to the first aspect of the invention. figure 7 represents an example of an aging curve. figure 8 represents the relative state of the aircraft engine over time. DESCRIPTION DETAILLEE
[0031] Unless otherwise specified, the same element appearing on different figures has a unique reference.
[0032] A first aspect of the invention relates to a method for evaluating the condition of an aircraft engine.
[0033] The aircraft is, for example, an airplane, a helicopter, or a drone.
[0034] The engine condition is assessed for a given flight of the aircraft, that is to say, the condition of the aircraft engine is considered not to vary during a flight.
[0035] In order to implement the process according to the first aspect of the invention, the aircraft must have carried out a plurality of maintenances and a plurality of flights between two successive maintenances, before the flight for which the condition of the engine is evaluated.
[0036] [ Fig. 3 ] There figure 3 illustrates an example of previous maintenance M i and flights V i,j of the aircraft for which the engine condition is evaluated for a given flight V.
[0037] To an i-th maintenance M i is associated a set of flights V i,j , the notation flight V i,j corresponds to the j-th flight which took place between the maintenance M i-1 and the maintenance M i .
[0038] Each maintenance M i is associated with a maintenance instant t Mi, corresponding for example to the beginning or end of the time interval required to carry out the maintenance M i.
[0039] On the figure 3 , the aircraft carried out three maintenance M 0 , M 1 , M 2 associated respectively with the maintenance times t M0 , t M1 , t M2 .
[0040] The first maintenance M 0 of the aircraft is considered to be carried out before it enters service, i.e. before its first flight V 1.0.
[0041] During each flight Vi,j, V, data are recorded periodically. More specifically, sets Xi,j,k of variables, each associated with a recording time tEi,j,k, are recorded. The sets Xi,j,k of variables can be recorded at a fixed time interval, meaning that the time interval between the recording times tEi,j,k, tEi,j,k+1 of two successively recorded sets of variables Xi,j,k and Xi,j,k+1 is fixed, or at a variable time interval.
[0042] The sets X i,j,k of variables recorded for the same flight V i,j , V or for different flights V i,j , V all have the same variables.
[0043] Each set of variables includes, for example, at least one variable related to the engine, such as torque, temperature, or pressure. One or more environmental variables can be added, such as outside temperature, atmospheric pressure, flight conditions, and / or one or more flight parameters, such as flight controls or onboard payload.
[0044] Each flight V i,j , V is associated with a time series S i,j , S comprising the sets X i,j,k of variables recorded successively during the flight in question. The sets X i,j,k of variables are ordered in the order of recording, that is to say that a first set X i,j,k of variables recorded at a first recording time t Ei,j,k prior to a second recording time t Ei,j,k+1 of a second set X i,j,k+1 of variables is located before this second set X i,j,k+1 of variables in the time series S i,j , S.
[0045] Each time series S i,j , S is associated with a flight time t Vi,j , tv corresponding for example to the beginning or end of the flight Vi,j , V corresponding to the time series S i,j , S.
[0046] On the figure 3 The aircraft performed five flights V1,0, V1,1, V2,0, V2,1, V2,2, associated respectively with the time series S1,0, S1,1, S2,0, S2,1, S2,2, before the flight V under consideration; two flights V1,0, V1,1 between the first maintenance M0 and the second maintenance M1; and three flights V2,0, V2,1, V2,2 between the second maintenance M1 and the third maintenance M2. The time series S1,0, S1,1, S2,0, S2,1, S2,2 are respectively associated with the flight instants tV1,0, tV1,1, tV2,0, tV2,1, tv2,2.
[0047] On the figure 3 is represented for flight V 2,0 , the sets of variables X 2,0,0 , X 2,0,1 , X 2,0,2 , X 2,0,3 , X 2,0,4 , X 2,0,5 , X 2,0,6 , X 2,0,7 recorded respectively at the recording times t E2,0,0 , t E2,0,1 , t E2,0,2 , t E2,0,3 , t E2,0,4 , t E2,0,5 , t E2,0,6 , t E2,0,7 . The time series S 2,0 therefore includes the sets of variables X 2,0,0 , X 2,0,1 , X 2,0,2 , X 2,0,3 , X 2,0,4 , X 2,0,5 , X 2,0,6 , X 2,0,7 ordered according to the order of the recording times t E2,0,0 , t E2,0,1 , t E2,0,2 , t E2,0,3 , t E2,0,4 , t E2,0,5 , t E2,0,6 , t E2,0,7 .
[0048] The state assessed by the method according to the first aspect of the invention for the flight V under consideration is relative because it is obtained by comparison with a reference flight selected from among the flights Vi,j previously performed by the aircraft. The time series Si,j associated with the reference flight is called the reference time series.
[0049] [ Fig. 1 ] There figure 1 is a synoptic diagram illustrating the sequence of steps of process 100 according to the first aspect of the invention.
[0050] A first step 101 of process 100 consists of creating a dataset Ji for each maintenance M i performed by the aircraft before the flight V under consideration.
[0051] The dataset Ji includes the time series S i,j associated with each flight V i,j that was carried out between the maintenance time t Mi associated with the maintenance M i and the maintenance time t Mi-1 associated with the maintenance M i-1 immediately preceding the maintenance M i.
[0052] Since no flights were conducted prior to the first maintenance M 0, no dataset J 0 is created for the first maintenance M 0.
[0053] On the figure 3 A dataset J1 is created for maintenance M1 containing the time series S1,0 and S1,1 associated respectively with the flights V1,0 and V1,1 which took place between the first maintenance M0 and maintenance M1. A dataset J2 is created for maintenance M2 containing the time series S2,0, S2,1, S2,2 associated respectively with the flights V2,0, V2,1, V2,2 which took place between maintenance M1 and maintenance M2.
[0054] A second step 102 of process 100 is to create a set of datasets comprising each dataset Ji created in the first step 101.
[0055] The dataset set includes at least one Ji dataset.
[0056] Taking the example of the figure 3 The dataset set comprises two datasets, J1 and J2.
[0057] The second step, 102, of process 100 then consists of separating the dataset into a training set and a validation set. For example, if the dataset contains N datasets, the training set contains n datasets from among the N datasets in the dataset, with n strictly less than N, and the validation set contains Nn datasets. In general, and depending on the number of datasets available, n is chosen so that the training set contains more datasets than the validation set.
[0058] If the dataset set comprises a single dataset, the training set and the validation set each include, for example, a dataset comprising a part of the single dataset in the dataset set.
[0059] A third step 103 of process 100 consists of jointly training an embedding function and a classifier.
[0060] [ Fig. 4 ] There figure 4 shows a schematic representation of the third step 103 of process 100.
[0061] The embedding function 201 is a parametric function configured to associate a point, called the embedding Pi,j, in an embedding space to a time series Si,j. The embedding function 201 is chosen based on the space of time series Si,j, S.
[0062] The embedding function 201 is for example a neural network, a density distribution, a parametric analytic function or a representation function defining a Hilbert reproducing kernel.
[0063] The 202 classifier is configured to associate with a pair of points in the embedding space, that is, with a pair of embeddings Pi,j comprising a first embedding Pi,j1 and a second embedding Pi,j2, a scalar between 0 and 1 corresponding to a probability Pi,j1,i,j2. The 202 classifier is chosen according to the embedding space.
[0064] The 202 classifier is, for example, a linear classifier, such as a perceptron, a support vector machine or a Bayesian classifier, a quadratic classifier, a neural network or a decision tree.
[0065] The training is carried out by minimizing a cost function H depending on the probability pi,j1,i,j2, allowing on the one hand to update the parameters of the classifier 202 so that the probability pi,j1,i,j2 calculated by the classifier 202 corresponds to the probability that a first time series S i,j1 corresponding to the first dive P i,j1 is associated with a flight time t Vi,j1 prior to the flight time t Vi,j2 associated with a second time series S i,j2 corresponding to the second dive P i,j2, and on the other hand to update the parameters of the dive function 201 so that it is able to order the dives P i,j corresponding to the time series S i,j according to the temporal order in the dive space.
[0066] The third step 103 includes several sub-steps performed for each dataset Ji of the training set.
[0067] A first sub-step 1031 of the third step 103 consists of randomly drawing from the dataset Ji of the training set, a pair of time series comprising a first time series S i,j1 and a second time series S i,j2.
[0068] On the figure 4 , the first substep 1031 is carried out for a dataset Ji comprising three flights Vi,j1, Vi,j2, Vi,j3 associated respectively with the time series Si,j1, Si,j2, Si,j3, and the time series Si,j1 and Si,j2 are drawn randomly.
[0069] A second substep 1032 of the third step 103 consists of using the embedding function 201 to associate the pair of time series drawn in the first substep 1031 with a pair of embeddings Pi,j1, Pi,j2.
[0070] On the figure 4 , an embedding P i,j1 is obtained for the time series S i,j1 and an embedding P i,j2 is obtained for the time series S i,j2 using the embedding function 201.
[0071] A third substep 1033 of the third step 103 consists of using the classifier 202 to obtain the probability pi,j1,i,j2 for the pair of embeddings P i,j1 , P i,j2 obtained in the second substep 1032.
[0072] On the figure 4 , a probability pi,j1,i,j2 is obtained for a pair of embeddings including the embedding P i,j1 and the embedding P i,j2 using the 202 classifier.
[0073] A fourth sub-step 1034 of the third step 103 consists of calculating the cost function H from the probability pi,j1,i,j2 obtained in the third sub-step 1033, the time of flight t Vi,j1 corresponding to the first time series S i,j1 and the time of flight t Vi,j2 corresponding to the second time series S i,j2.
[0074] The cost function H is, for example, a binary entropy function, expressed as: H p i , j 1 , i , j 2 = c t Vi , j 1 t Vi , j 2 log p i , j 1 , i , j 2 + 1 − c t Vi , j 1 t Vi , j 2 log 1 − p i , j 1 , i , j 2
[0075] With c equal to 0 if the time of flight t Vi,j1 corresponding to the first time series S i,j1 is later than the time of flight t Vi,j2 corresponding to the second time series S i,j2 and equal to 1 if the time of flight t Vi,j1 corresponding to the first time series S i,j1 is earlier than the time of flight t Vi,j2 corresponding to the second time series S i,j2.
[0076] In the case of the figure 4 , c equals 1 since the first time series S i,j1 is associated with a flight time t Vi,j1 prior to the flight time t Vi,j2 corresponding to the second time series S i,j2 .
[0077] The cost function H is, in another example, a regularized binary entropy function, expressed as: H p i , j 1 , i , j 2 = c t Vi , j 1 t Vi , j 2 log p i , j 1 , i , j 2 + 1 − c t Vi , j 1 t Vi , j 2 log 1 − p i , j 1 , i , j 2 + ‖ P i , j 1 ‖ + ‖ P i , j 2 ‖
[0078] With ∥.∥ corresponding to the regularization term, for example by the L1 norm ∥.∥ 1 to obtain a sparse embedding or the L2 norm ∥.∥ 2 to obtain a spherical embedding centered at 0.
[0079] The cost function H is, in another example, the following function: H p i , j 1 , i , j 2 = ‖ c t Vi , j 1 t Vi , j 2 − p i , j 1 , i , j 2 ‖
[0080] A fifth substep 1035 of the third step 103 consists of optimizing the embedding function 201 and the classifier 202 by minimizing the cost function H calculated in the fourth substep 1034.
[0081] [ Fig. 5 ] There figure 5 shows a schematic representation of a fourth and fifth step 105 of process 100.
[0082] The fourth step 104 of process 100 consists of selecting the reference time series S ref from among the time series S i,j included in the validation set.
[0083] [ Fig. 2 ] There figure 2 is a synoptic diagram illustrating the sequence of sub-steps for a first embodiment of the fourth step 104 of process 100.
[0084] [ Fig. 6 ] There figure 6 shows a schematic representation of a first substep 1041, a second substep 1042 and a third substep 1043 of the first embodiment of the fourth step 104 of process 100.
[0085] The first substep 1041 of the fourth step 104 according to the first embodiment consists of applying the trained embedding function 201 to each time series S i,j of the validation set, to obtain an embedding P i,j for each time series S i,j of the validation set.
[0086] On the figure 6 The validation set comprises a first dataset Ji including two time series S i,j1 , S i,j2 and a second dataset J l including three time series S l,j1 , S l,j2 , S l,j3 . During the first substep 1041 of the fourth step 104 according to the first embodiment, an embedding P i,j1 is obtained for the time series S i,j1 , an embedding P i,j2 is obtained for the time series S i,j2 , an embedding P l,j1 is obtained for the time series S l,j1 , an embedding P l,j2 is obtained for the time series S l,j2 and an embedding P l,j3 is obtained for the time series S l,j3 .
[0087] The second substep 1042, the third substep 1043, the fourth substep 1044 and the fifth substep 1045 are performed for each time series S i,j of the validation set.
[0088] The second substep 1042 of the fourth step 104 according to the first embodiment consists of using the embedding P i,j calculated in the first substep 1041 for each time series S i,j of the validation set, to calculate a distance d between the embedding P i,j of a time series S i,j considered and the embedding of each of the other time series of the validation set.
[0089] For example, on the figure 6 The time series under consideration is the time series Si,j1, which is represented by a bounding box. During the second substep 1042 of the fourth step 104, according to the first embodiment, a distance d(Pi,j1, Pi,j2) is calculated between the time series Si,j1 and the time series Si,j2, a distance d(Pi,j1, Pl,j1) is calculated between the time series Si,j1 and the time series Sl,j1, a distance d(Pi,j1, Pl,j2) is calculated between the time series Si,j1 and the time series Sl,j2, and a distance d(Pi,j1, Pl,j3) is calculated between the time series Si,j1 and the time series Sl,j3. The distance d is a distance on the embedding space. For example, if the embedding space is a distribution space, the distance d can be any distance defined on the distribution space.If the embedding space is a vector space, the distance d can be any distance defined on a vector space. A distance d (P i,j1 , P l,j2 ) calculated between the vector embedding P i,j1 of the time series S i,j1 and the vector embedding P l,j2 of the time series S l,j2 is for example the distance L2 between the vectors P i,j1 and P l,j2 which is expressed as follows: . ‖ P i , j 1 − P l , j 2 ‖
[0090] The third substep 1043 of the fourth step 104 according to the first embodiment consists of creating an aging curve C i,Si,j for each dataset Ji of the validation set and each time series S i,j of the validation set under consideration.
[0091] [ Fig. 7 ] There figure 7 represents an example of an aging curve for a dataset.
[0092] For a given dataset J l, the aging curve C l,Si,j associates to each instant of flight t Vl,j associated with a time series S l,j of the validation set included in the dataset J l, the distance d calculated in the second sub-step 1042 of the fourth step 104 according to the first embodiment, between the time series S i,j considered and the time series S l,j of the validation set included in the dataset J l.
[0093] On the figure 6 The validation set consists of only two datasets Ji and Jl. During the third substep 1043 of the fourth step 104 according to the first embodiment, an aging curve Ci,Si,j1 is created for the dataset Ji and an aging curve Cl,Si,j1 is created for the dataset Ji. The time series under consideration being the time series S i,j1, the aging curve C i,Si,j1 includes the distance d (P i,j1 , P i,j2 ) between the time series S i,j1 under consideration and the time series S i,j2 , and the aging curve C l,Si,j1 includes the distance d (P i,j1 , P l,j1 ) between the time series S i,j1 under consideration and the time series S l,j1 , the distance d (P i,j1 , P l,j2 ) between the time series S i,j1 under consideration and the time series S l,j2 and the distance d (P i,j1 , P l,j3 ) between the time series S i,j1 under consideration and the time series S l,j3 .
[0094] On the figure 7 The aging curve Ci, Si,j1 calculated for the dataset Jl with respect to the considered time series Si,j is represented as a function of time. The aging curve Ci, Si,j1 is, for example, a continuous function with respect to time constructed by linear interpolation.
[0095] The fourth substep 1044 of the fourth step 104 according to the first embodiment consists of calculating a monotonicity indicator for each aging curve C i,Si,j created in the third substep 1043 of the fourth step 104 according to the first embodiment.
[0096] The fourth substep 1044 of the fourth step 104 according to the first embodiment includes for example a first substep consisting of initializing the monotony indicator to zero, then a second substep consisting of calculating the first-order time derivative of the aging curve C i,Si,j for each time interval between two successive flight instants t Vi,j and incrementing the monotony indicator by 1 when the calculated derivative is positive.
[0097] The monotonicity indicator of an aging curve then corresponds to the number of time intervals between two successive flight instants t Vi,j for which the first-order time derivative of the aging curve C i,Si,j is positive.
[0098] The fifth substep 1045 of the fourth step 104 according to the first embodiment consists of calculating an aging indicator from each monotony indicator calculated in the fourth substep 1044 of the fourth step 104 according to the first embodiment.
[0099] The aging indicator is, for example, the sum of the monotony indicators calculated in the fourth substep 1044 of the fourth step 104 according to the first embodiment.
[0100] At the end of the fifth sub-step 1045 of the fourth step 104 according to the first embodiment, an aging indicator is calculated for each time series S i,j of the validation set.
[0101] The reference time series S ref is then the time series S i,j for which the aging indicator calculated in the fifth sub-step 1045 of the fourth step 104 according to the first embodiment meets an aging condition.
[0102] The aging condition is checked for example for the maximum aging indicator.
[0103] The embedding P i,j calculated at the first substep 1041 of the fourth step 104 according to the first embodiment for the reference time series S ref is called the reference embedding P ref.
[0104] According to the second embodiment, the reference time series S ref is the time series S 1,0 associated with the first flight V 1,0 of the aircraft.
[0105] The fourth step 104 according to the second embodiment includes a substep of using the plunge function 201 driven on the reference series S ref to obtain a reference plunge P ref.
[0106] The fifth step 105 of process 100 consists of calculating an engine status indicator for the flight V under consideration.
[0107] A first substep 1051 of the fifth step 105 consists of applying the embedding function 201 trained on the time series S associated with flight V to obtain an embedding P, as illustrated in the figure 5 .
[0108] A second substep 1052 of the fifth step 105 consists of calculating the distance d between the immersion P calculated in the first substep 1051 of the fifth step 105 and the reference immersion P ref , as illustrated in the figure 5 .
[0109] The engine status indicator is then equal to the distance d calculated in the second sub-step 1052 of the fifth step 105.
[0110] [ Fig. 8 ] There figure 8 This represents the relative state of the aircraft engine over time, in the form of a curve that represents the engine condition indicator as a function of time. The vertical lines represent the times of the maintenance performed. This curve thus exhibits a discontinuity at each maintenance Mi; it is approximately monotonic between two maintenance Mi, corresponding to the aging curve of a dataset calculated between these two maintenances.
[0111] The process 100 may further include a sixth step 106 of triggering an alarm if the engine status indicator is above an alarm threshold SA.
[0112] The SA alarm threshold is, for example, equal to the maximum aging indicator preceding a maintenance M i of a predefined time margin, that is to say, the aging indicator with the highest value among a set of aging indicators including each aging indicator calculated at a time preceding a maintenance M i of the time margin.
[0113] A second aspect of the invention relates to a method for monitoring the relative state of the aircraft engine during a set of flights V.
[0114] The monitoring method according to the second aspect of the invention includes, for example, the steps of the evaluation method 100 according to the first aspect of the invention for a first flight V of the flight set, then the sixth step 106 of the evaluation method 100 according to the first aspect of the invention for the other flights V of the flight set.
[0115] A third aspect of the invention relates to a computer program product comprising instructions which, when the program is executed on a computer, lead the computer to implement the steps of the evaluation process according to the first aspect of the invention or the steps of the monitoring process according to the second aspect of the invention.
[0116] The computer program product according to the invention is a surveillance or monitoring tool, preferably implemented on the ground.
Claims
1. A method (100) for evaluating the relative state of an aircraft engine for a given flight (V) of the aircraft, the aircraft having performed a plurality of maintenance operations (Mi) prior to the given flight (V) and a plurality of flights (Vi,j) between two successive maintenance operations (Mi), each maintenance operation (Mi) being associated with a maintenance time instant (tMi), a plurality of collections (Xi,j,k) of variables being recorded during each flight (Vi,j, V) in a recording order, each flight (Vi,j, V) being associated with a time-domain series (S, Si,j) corresponding to a flight time instant (tv, tVi,j) and including the plurality of collections (Xi,j,k) of variables recorded during the flight (Vi,j, V) ordered according to the recording order, the method (100) being characterised in that it includes the following steps of: - For each maintenance operation (Mi), creating a dataset (Ji) including the time-domain series (Si,j) associated with each flight (Vi,j) that has been carried out between the maintenance time instant (tMi) associated with the maintenance operation (Mi) and the maintenance time instant (tMi-1) associated with the previous maintenance operation (Mi-1, 101); - Creating a collection of datasets comprising each dataset (Ji) created and separating the collection of datasets (Ji) into a training collection and a validation collection (102); - Jointly training an embedding function (201) and a classifier (202) by minimising a cost function (H, 103), including the following sub-steps of: ∘ For each dataset (Ji) of the training collection: • Randomly drawing a pair of time-domain series including a first time-domain series (Si,j1) and a second time-domain series (Si,j2, 1031); • Applying the embedding function (201) to the pair of time-domain series to obtain a pair of embedments (Pi,j1, Pi,j2, 1032); • Using the classifier (202) on the pair of embedments (Pi,j1, Pi,j2) to obtain a probability (pi,j1,i,j2) that the first time-domain series (Si,j1) corresponds to a flight time instant (tVi,j1) earlier than the flight time instant (tVi,j2) corresponding to the second time-domain series (Si,j2, 1033); • Calculating the cost function (H) from the probability (pi,j1,i,j2) obtained, the flight time instant (tVi,j1) corresponding to the first time-domain series (Si,j1) and the flight time instant (tVi,j2) corresponding to the second time-domain series (Si,j2, 1034); • Optimising the embedding function (201) and the classifier (202) by minimising the cost function (H) calculated (1035); - Selecting a reference time-domain series (Sref) from the time-domain series (Si,j) included in the validation collection, including a sub-step of applying the trained embedding function (201) to the reference time-domain series (Sref) to obtain a reference embedment (Pref, 104); - Calculating an indicator of engine state for the given flight (V) (105), including the following sub-steps of: ∘ Applying the trained embedding function (201) to the time-domain series (S) associated with the flight (V), to obtain an embedment (P, 1051); ∘ Calculating a distance (d) between the embedment (P) and the reference embedment (Pref, 1052), the distance (d) calculated corresponding to the indicator of engine state.
2. The evaluation method (100) according to claim 1, characterised in that the step (104) of selecting the reference time-domain series (Sref) includes the following sub-steps of: - Applying the trained embedding function (201) to each time-domain series (Si,j) of the validation collection, to obtain an embedment (Pi,j) for each time-domain series (Si,j) of the validation collection (1041); - For each time-domain series (Si,j) of the validation collection : ∘ Calculating the distance (d) between the embedment (Pi,j) of the time-domain series (Si,j) and the embedment (Pi,j) of each other time-domain series (Si,j) of the validation collection (1042); ∘ For each dataset (Ji) of the validation collection, creating an ageing curve (Ci,Si,j) associating with each flight time instant (tVi,j) associated with a time-domain series (Si,j) of the validation collection included in the dataset (Ji), the corresponding distance (d) calculated (1043); ∘ For each ageing curve (Ci,Si,j), calculating a monotonicity indicator (1044); ∘ Calculating an ageing indicator from the monotonicity indicators calculated (1045); the reference time-domain series (Sref) being the time-domain series (Si,j) for which the ageing indicator calculated meets an ageing condition.
3. The evaluation method (100) according to claim 2, characterised in that the sub-step (1044) of calculating the monotonicity indicator includes the following sub-steps of: - Initialising the monotonicity indicator to zero; - For each time interval between two successive flight time instants (tVi,j), calculating the first-order time derivative of the ageing curve (Ci,Si,j): ∘ If the derivative is positive, incrementing the monotonicity indicator by 1.
4. The evaluation method (100) according to claim 3, characterised in that the ageing indicator is the sum of the monotonicity indicators calculated.
5. The evaluation method (100) according to claim 4, characterised in that the ageing condition is verified for the maximum ageing indicator calculated.
6. The evaluation method (100) according to any of the preceding claims, characterised in that the cost function (H) is a binary entropy function.
7. The evaluation method (100) according to any of the preceding claims, characterised in that it further includes a step (106) of triggering an alarm if the indicator of engine state is greater than an alarm threshold (AT).
8. The evaluation method (100) according to claims 2 and 7, characterised in that the alarm threshold (AT) is equal to the maximum ageing indicator preceding a maintenance operation (Mi) by a predefined time margin.
9. A method for tracking the relative state of an aircraft engine over a plurality of flights (V) including the steps of the evaluation method (100) according to any of the preceding claims for one flight (V) of the plurality of flights (V), and the triggering step (106) for every other flight (V) of the plurality of flights (V).
10. A computer program product comprising instructions which, when the program is executed on a computer, cause the computer to implement the steps of the evaluation method (100) according to any one of claims 1 to 8 or the steps of the tracking method according to claim 9.
Citation Information
Patent Citations
Computing environment system for monitoring aircraft engines
WO2020115440A1
DETECTION OF ANOMALY IN AN AIRCRAFT ENGINE.
FR2939170A1
METHODS FOR CREATING A DATABASE AND DEVELOPING A STATE CHART FOR AIRCRAFT ENGINES, AND A METHOD FOR MONITORING THE OPERATION OF AN ASSOCIATED AIRCRAFT ENGINE
FR3006785A1
ENGINE HEALTH MONITORING SYSTEM AND ASSOCIATED CONFIGURATION METHOD
FR3035232A1
